README.md

---
language:
- en
- zh
license: apache-2.0
library_name: transformers
tags:
- chat-template
- reasoning
- cot
- protium
- solstice-ai
- objective-gated
- zero-waste
- dynamic-reasoning
- gguf
- ollama
- llama.cpp
- vllm
---

<p align="center">
  <img src="https://cdn-uploads.huggingface.co/production/uploads/67c2e844e0921a5410eec10a/Y5M42dCag2f7Fc6fDtV0Z.jpeg" alt="Solstice-AI Banner" width="100%">
</p>

<h1 align="center">Protium-1.0: Objective-Gated Cognitive Architecture</h1>

<h3 align="center">Zero-Waste Cognitive Protocol &bull; Dynamic Cognitive Impedance &bull; Universal Cross-Family Architecture</h3>

<p align="center">
  <img src="https://img.shields.io/badge/org-Solstice--AI-blueviolet" alt="Solstice-AI">
  <img src="https://img.shields.io/badge/license-Apache%202.0-blue" alt="License">
  <img src="https://img.shields.io/badge/protocol-Protium--1.0%20(Objective--Gated)-success" alt="Protocol">
  <img src="https://img.shields.io/badge/architecture-Universal%20Cross--Family-crimson" alt="Architecture">
  <img src="https://img.shields.io/badge/default%20mode-L10%20Proteus%20Dynamic-purple" alt="Default Mode">
  <img src="https://img.shields.io/badge/inference-Zero--Waste%20Early--Exit-brightgreen" alt="Zero-Waste">
</p>

---

## Executive Summary

**Protium-1.0** is Solstice-AI's high-efficiency, objective-gated cognitive reasoning architecture for open-weight foundation models. Named after **Protium** ($^1\text{H}$) — the fundamental isotope of hydrogen featuring zero neutrons and zero dead weight — Protium-1.0 replaces traditional, unconstrained chain-of-thought bloat with a lean, forward-only reasoning pipeline.

Rather than imposing static token minimums, rigid pseudo-schemas, or unconstrained internal deliberation loops, Protium-1.0 operates on an **Objective-Gated Protocol**: every reasoning sequence anchors its verification criteria at step one, reasons strictly forward toward that objective, and terminates deliberation the moment verified correctness is attained.

Equipped with the **`[L10]` Proteus Dynamic Engine** as its native default, Protium-1.0 auto-calibrates cognitive depth to the intrinsic difficulty of each prompt—minimizing overhead on direct queries while dynamically allocating full analytical runway for axiomatic proofs and complex engineering tasks.

---

## Architectural Principles

* **Token 1 Target Anchoring**: At the inception of the cognitive trajectory, the model isolates and locks the exact termination criteria required for a complete, verified solution.
* **Forward-Only Deduction**: Reasoning progresses strictly forward along deductive axes, eliminating internal conversational stalling and rhetorical hesitation from the attention cache.
* **Objective-Gated Early Exit**: Deliberation is bounded exclusively by mathematical, programmatic, and empirical verification. The moment the target anchor is verified, thinking terminates immediately and the final response is emitted.
* **Delta-Isolated Error Correction**: If a logical contradiction or boundary failure is detected during verification, the model isolates and repairs only the specific broken parameter, preventing circular loops or full-context restarts.
* **Decisive Operational Delivery**: Outputs are delivered with high information density, clear technical rigor, and zero conversational filler.

---

## The Cognitive Spectrum

Protium-1.0 features 11 calibrated reasoning tiers plus specialized multi-agent convergence modes. It defaults natively to **`[L10]` Proteus (Dynamic Adaptive)**:

| Level | Identifier | Archetype | Operational Focus | Workload Profile |
|---|---|---|---|---|
| **[L10]** | `proteus` / `dynamic` | **Dynamic Adaptive (Default)** | **Fluid impedance matching** | **Auto-calibrated depth across all inputs; collapses on lookups, expands on proofs** |
| **[L0]** | `mortal` / `regular` | Instant Execution | Thinking Suppressed | Conversational dialogues, zero-overhead classification, streaming tasks |
| **[L1]** | `hermes` | Rapid Instinct | High-velocity verification | Single-pass sanity checks, syntax validation, arithmetic confirmation |
| **[L2]** | `apollo` | Disciplined Deduction | Linear constraint trace | Parameter mapping, rule following, structured deduction |
| **[L3]** | `artemis` | Boundary & Failure Audit | Edge vulnerability probe | Null/empty states, numeric limits, concurrency collision auditing |
| **[L4]** | `athena` | Strategic Architecture | Modular synthesis | Multi-step refactoring, API interface design, trade-off optimization |
| **[L5]** | `prometheus` | Defensive Engineering | Second-order impact modeling | Fault tolerance, graceful degradation, scaling lifecycle resilience |
| **[L6]** | `solstice` | Deep Systemic Derivation | Multi-hypothesis synthesis | Axiomatic proofs, competing architecture stress-testing |
| **[L7]** | `hyperion` | First-Principles Synthesis | Ground-up decomposition | Comprehensive first-principles decomposition across all logical surfaces |
| **[L8]** | `einstein` | Theoretical Mastery | Unified invariant reduction | Theoretical breakthroughs, advanced algorithmic reduction, formal derivations |
| **[L9]** | `oracle` | Proof-Grade Authority | Formal mathematical verification | Exhaustive formal proofs, zero-placeholder production implementations |
| **Swarm** | `swarm` | 20-Perspective Swarm | Divergent conceptual synthesis | Complex multi-perspective cross-pollination and convergence |
| **Council**| `council` | Adversarial Research Panel | Institutional consensus | Formal verification critic panel, candidate scoring, and audit |

---

## Steering & Control Reference

Protium-1.0 supports flexible steering via inline tags or standard inference parameters:

| Control Target | Inline Syntax | API Kwarg / Parameter |
|---|---|---|
| **Dynamic Adaptive (Default)** | `[L10]` or `{REASON:proteus}` | `reasoning_effort="proteus"` or `"dynamic"` |
| **Specific Cognitive Levels** | `[L1]` through `[L9]` | `reasoning_effort="1"` .. `"9"` (or level name) |
| **Silent Reasoning** | `{REASON:i8}` | Runs high-tier analytical rigor with `<think>` hidden |
| **Zero-Overhead Vanilla** | `[regular]` or `[vanilla]` | `regular=True` or `mode="vanilla"` |

---

## Template Family Distribution

All templates are located in the `templates/` directory and pre-configured for native runtime token formats:

| Target Model Architecture | File | Native Format |
|---|---|---|
| **Universal / ChatML** | `templates/universal.jinja` | `<|im_start|>`, `<|im_end|>` |
| **Qwen 2.5 / 3.8 / Vision** | `templates/qwen.jinja` | `<|im_start|>`, `<|im_end|>`, Multimodal Vision |
| **DeepSeek Native** | `templates/deepseek.jinja` | `<｜User｜>`, `<｜Assistant｜>`, Native DSML |
| **Meta Llama 3 / 4** | `templates/llama3.jinja` | `<|start_header_id|>`, `<|eot_id|>` |
| **Mistral / Mixtral** | `templates/mistral.jinja` | `[INST]`, `[/INST]` |
| **Google Gemma 2 / 3** | `templates/gemma.jinja` | `<start_of_turn>`, `<end_of_turn>` |
| **Zhipu GLM 4 / 5.3** | `templates/glm.jinja` | `[gMASK]<sop>`, `<|user|>`, `<|assistant|>` |
| **Vanilla Reference** | `templates/regular_*.jinja` | Minimal, zero-overhead baseline templates |

---

<div align="center">
Developed and maintained by <b>Solstice AI</b>.
</div>